In recent years, non-destructive imaging technology has seen rapid advancements in the fields of painting research and conservation. Among these, macro X-ray fluorescence (MA-XRF) analysis stands out, aiding experts in identifying pigments and analyzing painting techniques, thereby providing valuable insights into the artistic creation process. However, the MA-XRF technique generates large and complex datasets, posing challenges to traditional data analysis methods.
Recently, Italian researchers have applied deep learning algorithms to spectral analysis of MA-XRF datasets, developing a novel analytical approach. This method uses over 500,000 synthetic spectra generated by Monte Carlo simulations to train the deep learning algorithm, enabling rapid and accurate analysis of XRF spectra in MA-XRF datasets, overcoming the limitations of traditional deconvolution methods.
To validate the accuracy and applicability of the new method, researchers applied it to two Raphael paintings, "God the Father" and "Madonna," exhibited at the Capodimonte Museum in Italy. The results showed that the deep learning model not only accurately quantified fluorescence line intensities but also effectively eliminated artifacts produced by traditional analysis methods, generating clearer element distribution maps.
Through comparative analysis with traditional deconvolution algorithms, researchers found that the new method performs better in handling elements with low counts and low signal-to-noise ratios, accurately separating overlapping fluorescence lines in XRF spectra, thus more precisely identifying pigments. For example, the new method can accurately distinguish between elements with similar energies, such as iron (Fe) and manganese (Mn), as well as lead (Pb) and sulfur (S), avoiding misjudgments common in traditional methods.
This research marks a significant advancement in the application of artificial intelligence technology in the analysis of art works, providing new insights for more accurate and efficient analysis of XRF spectra, especially in dealing with large datasets generated by MA-XRF imaging technology. In the future, researchers plan to further expand the application scope of this method, such as inferring the stratification structure of paintings or comparing spectral data obtained from different instruments.
Paper link: https://www.science.org/doi/10.1126/sciadv.adp6234